An effective approach to preventing TCP incast throughput collapse for data center networks

This paper presents an effective solution to the known TCP incast problem in data center networks. The incast problem refers to a drastic TCP throughput drop when the number of servers synchronically sending data to the same receiver is too large. Our proposed approach utilizes the link bandwidth as fully as possible but without significant packet losses by limiting the number of concurrent senders to a reasonable value. Based on the concept of bandwidth delay product, our approach conservatively estimates the reasonable number of concurrent senders. Our approach does not modify TCP protocol itself and can thus be applied to any TCP variant, and works regardless of the type of data center network topology and throughput limitation. Analysis and simulation results show that our approach eliminates the incast problem and noticeably improves TCP throughput.